New Report Warns Policy Gaps Could Stifle Innovation in Open AI Foundation Models

The absence of clear policies for open AI models risks stifling innovation, according to a new report highlighting regulatory gaps and challenges.

In a newly published report “Considerations for governing open foundation models” in Science, a group of researchers led by Rishi Bommasani has highlighted the potential consequences of regulating open foundation models, stressing that existing policies may unintentionally hinder AI innovation. Foundation models, such as GPT-4 and Llama 2, represent the core technology behind many generative AI applications, and their influence on both technological progress and society has prompted widespread calls for regulation.

The debate surrounding the release of foundation models lies at the heart of current regulatory discussions. Open foundation models, such as Meta’s Llama 2, share their underlying parameters openly, while closed models, like OpenAI’s GPT-4, only offer restricted access through query interfaces. This difference is significant in determining who can develop and benefit from these technologies. Bommasani and colleagues argue that a balanced approach is crucial to foster innovation while managing associated risks.

The report raises concerns about how different policies might unevenly affect open versus closed models, with open foundation models often being overlooked by current regulatory frameworks. The authors warn that without explicit consideration for open models, these regulations could disproportionately harm the innovation ecosystem surrounding them. Bommasani et al. recommend that policymakers take into account the concept of “marginal risk”—assessing how open models contribute to increased risks compared to closed models or existing technologies—as a guiding principle when crafting regulations.

Benefits and Risks of Open Models

The authors of the report emphasise that open foundation models have several societal benefits, including promoting competition, fostering transparency, and distributing power. By making their model weights accessible, developers of open models allow broader research, enhance customisation, and support collaborative improvements—essentially lowering barriers to entry and ensuring that foundational AI advancements are not controlled solely by a few powerful entities.

However, the report also highlights the risks that come with open models. These include the potential misuse of the models for disinformation, cybersecurity threats, or malicious applications such as bioweapon development. The authors note that closed models might be better positioned to enforce restrictions and controls, but even closed systems remain vulnerable to some of these threats.

The researchers also draw an analogy between open foundation models and open-source software, noting that while both have distinct advantages in terms of innovation and accessibility, open foundation models differ due to the unique challenges associated with controlling how they are used. Despite these challenges, there is no concrete evidence suggesting that open foundation models are inherently more insecure or prone to misuse compared to their closed counterparts.

Impact on Researchers and Scientists

For scientists and researchers developing and using AI models in their labs, the implications of this report are significant. Open foundation models provide valuable opportunities for experimentation, customisation, and advancing the state of the art in AI research. The availability of model weights and underlying data fosters an environment where researchers can build upon existing work, test novel ideas, and create innovative applications without the financial and logistical barriers that often accompany closed models. This accessibility is crucial for smaller research labs and academic institutions that may not have the resources to access proprietary models.

However, the uncertainty surrounding regulatory frameworks presents a challenge. Researchers who rely on open models might face increasing compliance burdens or restrictions if policies are not designed with their needs in mind. This could potentially slow down research progress and limit the ability of scientists to contribute to the broader AI ecosystem. The report’s call for a balanced approach is particularly relevant for researchers who depend on open models to ensure that their work remains feasible and impactful.

The authors argue that fostering an open environment is essential for transparency and reproducibility in scientific research. Open models enable scientists to verify results, replicate studies, and advance knowledge in a collaborative manner. Policymakers must therefore consider the needs of the research community when crafting regulations to ensure that the benefits of open foundation models—such as promoting innovation, transparency, and broad participation—are not inadvertently stifled.

Recommendations for Policymakers

Bommasani and his team call on policymakers to ensure that regulatory efforts explicitly consider the nuances of open models. Current proposals often ignore the open/closed distinction, leading to compliance burdens that may unintentionally favour larger, well-resourced developers of closed models. The report advocates for a more nuanced approach, where regulatory efforts should not categorically penalise openness, particularly given the positive track record of open technologies in driving innovation and ensuring transparency.

Furthermore, the authors argue for a more precise definition of what constitutes an open foundation model in regulatory frameworks. They recommend direct consultation with the open foundation model community, urging that their input be taken into account to create policies that balance innovation with safety.

The discussion around foundation models and their governance is becoming increasingly critical as AI continues to expand its influence in society. This report is a good reminder that fostering innovation through openness must be considered alongside managing risks, and that maybe a nuanced approach to regulation will be necessary to ensure a thriving AI landscape.